Leaderboard Challenge 2026

The Hilti x Trimble SLAM Challenge 2026 placed particular emphasis on 360° visual-inertial sensing and floorplan-aware localization in active construction sites. Compared with previous editions, the benchmark highlighted a different set of difficulties: wide-FOV dual-fisheye imagery, textureless and repetitive indoor areas, and the need to align trajectories with building floorplan coordinates. The Challenge received over 1'800 submissions, coming from 22 teams in the Localizaton category and 62 teams in the SLAM category. Over 25 institutions from both academia and industry participated, with the final number potentially higher due to anonymous submissions. Results were announced at the IEEE ICRA 2026 workshop on Open Challenges in Robotics for Asset Inspection and Management.

We sincerely congratulate ETH-THU-Stuttgart team for their victory in the Localization category, and ACDC-K @ KAIST winning in the SLAM category. Additionally, we offer kudos to SnT-ARG @ University of Luxembourg and SPARO @ Inha University for securing 2nd place in Localization and SLAM categories respectively.

	Team	Score
1	ETH-THU-Stuttgart 3000 CHF	2,196.3
2	SnT-ARG - University of Luxembourg 1000 CHF	2,161.4
3	OmniRecon - ETH Zürich	2,041.3
4	CUFE - Cairo University	1,919.0
5	Map-It Ralph @ ETHZ - ETH Zürich	1,569.9
6	Undisclosed	1,505.6
7	Undisclosed	482.8
8	Trajectory Titans - University of Girona	239.4
9	Floorplan-aware - Independent	220.4
10	Borzo - HSLU	195.9
11	Monado's Basalt - TUM CVG	144.8
12	Sina - Sinalab	21.0

Top teams (with scores > 0.0) as per 15.05.2026.

	Team	Score
1	ACDC-K - Urban Robotics Lab, KAIST 3000 CHF	2,410.0
2	SPARO - Inha University 1000 CHF	2,393.9
3	Undisclosed	2,320.7
4	QQ - Imperial College London	2,308.8
5	SnT-ARG - University of Luxembourg	2,299.3
6	Mobile Robotics Lab (MRL) - TU Munich / ETH Zurich	2,292.6
7	OmniRecon - ETH Zürich	2,246.3
8	CUFE - Cairo University	2,206.3
9	Monado's Basalt - TUM CVG	2,195.4
10	SamsungPoland - Samsung Electronics Polska	2,157.9
11	Map-It Ralph @ ETHZ - ETH Zürich	2,155.5
12	Floorplan-aware - Independent	2,135.2
13	PanoAir - SUN YAT-SEN UNIVERSITY	2,078.0
14	Undisclosed	2,061.5
15	TU Korea	1,908.6
16	E.R.I.C. - Evolutionary Robotics Inc.	1,834.5
17	Riibotics	1,740.2
18	Gyeongnam - Gyeongnam University	1,668.3
19	Undisclosed	1,740.2
20	SLAM LEE	1,555.5

Top 20 teams as per 15.05.2026.

Results

In the Localization track, the winner combined a visual-inertial SLAM frontend with floorplan-based global anchoring. Their system used OKVIS2-X with both fisheye cameras and online extrinsic calibration, then incorporated floorplan priors generated by Z-FLoc. This approach demonstrated the value of using building-level geometric priors to reduce drift and express the estimated trajectory directly in the floorplan coordinate frame.

In the SLAM track, the winning team introduced ACDC-VSLAM, an adaptive visual-inertial SLAM system designed specifically for the dual-fisheye camera setup. The method addressed fisheye distortion through spherical equal-area feature selection, improved robustness in texture-poor areas with point and line features, and handled the non-overlapping front-rear camera configuration through adaptive cross-camera loop closure. The strongest SLAM results show that careful treatment of camera geometry and scene structure is essential for high performance on this year's benchmark.

Please note that the leaderboard contains only first 20 entries and/or those above the score of 0.0.

Read the corresponding academic publication

Score Computation

Submissions are evaluated with an exponential scoring metric, where higher scores indicate higher trajectory accuracy. Each pose in a run contributes to the score based on its error: a perfect estimate receives the maximum contribution, while larger errors receive progressively lower scores.

The scoring curve is calibrated so that an error of 0 m corresponds to 100 points, while an error of 10 m corresponds to 1 point. Each individual run score is capped at 100 points, and the final score is the sum of the scores across all evaluated runs.

For the SLAM task, 25 out of the 30 recordings are evaluated. The five Early Release recordings are excluded because their ground truth was made available, resulting in a maximum achievable SLAM score of 2500 points.

For the Localization task, the same Early Release recordings are excluded. In addition, floor_UG2-2025-12-02_run_1 is not evaluated because no floorplan is provided for it. Therefore, 24 recordings are evaluated, with a maximum achievable Localization score of 2400 points.